monitoreo de degradación forestal con imágenes spot4

8/9/2019 Monitoreo de Degradacin Forestal con imgenes SPOT4

1/13

Mapping forest degradation in the Eastern Amazon from SPOT 4 through

spectral mixture models

Carlos Souza Jr.a,b,*, Laurel Firestonea, Luciano Moreira Silvaa, Dar Robertsb

aInstituto do Homem e Meio Ambiente da Amazonia, Imazon, BrazilbDepartment of Geography, University of California at Santa Barbara, Santa Barbara, CA 93106, USA

Received 21 June 2002; received in revised form 30 July 2002; accepted 2 August 2002

Abstract

In this paper, we present a methodology to map classes of degraded forest in the Eastern Amazon. Forest degradation field data, available

in the literature, and 1-m resolution IKONOS image were linked with fraction images (vegetation, nonphotosynthetic vegetation (NPV), soil

and shade) derived from spectral mixture models applied to a Satellite Pour Lobservation de la Terre (SPOT) 4 multispectral image. The

forest degradation map was produced in two steps. First, we investigated the relationship between ground (i.e., field and IKONOS data) and

satellite scales by analyzing statistics and performing visual analyses of the field classes in terms of fraction values. This procedure allowed

us to define four classes of forest at the SPOT 4 image scale, which included: intact forest; logged forest (recent and older logged forests in

the field); degraded forest (heavily burned, heavily logged and burned forests in the field); and regeneration (old heavily logged and old

heavily burned forest in the field). Next, we used a decision tree classifier (DTC) to define a set of rules to separate the forest classes using the

fraction images. We classified 35% of the forest area (2097.3 km2) as intact forest. Logged forest accounted for 56% of the forest area and 9%

of the forest area was classified as degraded forest. The resultant forest degradation map showed good agreement (86% overall accuracy) with

areas of degraded forest visually interpreted from two IKONOS images. In addition, high correlation (R2 = 0.97) was observed between the

total live aboveground biomass of degraded forest classes (defined at the field scale) and the NPV fraction image. The NPV fraction also

improved our ability to mapping of old selectively logged forests.

D 2003 Published by Elsevier Inc.

Keywords: Mixture models; Forest degradation; Amazon; Nonphotosynthetic vegetation; Selective logging; Fire

1. Introduction

Selective logging and burning are considered the main

causes of forest degradation in the Amazon upland forests

(Nepstad et al., 1999). Numerous studies have measured the

impacts of logging and burning activities on species diver-

sity, carbon content, and economic value (e.g., Cochrane &

Shulze, 1999; Gerwing, 2002; Holdsworth & Uhl, 1997;

Johns, Barreto, & Uhl, 1996). While deforestation in the

Amazon has been mapped (INPE, 2000), the extent of forest

degradation is unknown. The total area of degraded forest,

estimated from field surveys, is far larger than the area

deforested each year. Annual deforestation rates range

between 11,000 and 29,000 km2, while it has been estimated

that up to 800015,000 km2 are logged, 80,000 km2 are

burned, and 38,000 km2 are fragmented or degraded each

year (Nepstad et al., 1999).

Spatial information on forest degradation would enhance

the effectiveness of planning development, commercial

activities, and conservation activities, as well as improve

local and global ecological models and carbon budget

estimates. For example, a forest degradation map could

be used to improve fire risk models (e.g., IPAM et al.,

1998), since degraded forests are highly fire-prone (Holds-

worth & Uhl, 1997) and should be included in fire risk

analysis. Forest law enforcement programs could also

benefit from the use of a forest degradation map. For

example, deforestation maps are combined with maps of

property lines to identify areas of unauthorized deforesta-

tion and deforestation in areas of permanent protection

(Firestone & Souza, 2002; Souza & Barreto, 2001). A

forest degradation map could also be produced and used

to identify forest plots that have not been properly managed

for timber harvesting.

0034-4257/$ - see front matterD 2003 Published by Elsevier Inc.

doi:10.1016/j.rse.2002.08.002

* Corresponding author. Department of Geography EH3601, Univer-

sity of California at Santa Barbara, Santa Barbara, CA 93106, USA.

www.elsevier.com/locate/rseRemote Sensing of Environment 87 (2003) 494506


2/13

Due to the vast area of dense forest with difficult access,

remote sensing is the most efficient and economically viable

means of mapping forests in the region. Unlike deforesta-

tion, which is easily identifiable through visual interpreta-

tion (INPE, 2000) or automated processing of multispectral

satellite images (Roberts, Batista, Pereira, Walker, & Nel-

son, 1998; Shimabukuro, Batista, Mello, Moreira, & Duarte,1998) forest degradation is more difficult to identify. There

is a range of degradation, and forest regeneration tends to

obscure the spectral/spatial signature of the degraded forest.

Additional complication arises from the complex spatial

arrangement of the land cover components (i.e., dead

vegetation, forest islands, bare soil) that makes up a de-

graded forest class (Fig. 1).

Prior studies have developed methods to map either

logged (Souza & Barreto, 2000; Stone & Lefebvre, 1998;

Watrin & Rocha, 1992) or burned forest in the Amazon

(Cochrane & Souza, 1998). However, no study has pro-

posed an all-encompassing classification scheme that can be

meaningfully linked with field studies in the region. Without

the ability to link mapped classes with data gathered in the

field, the usefulness of mapping methods is limited. In this

study, we evaluated the ability to link forest degradation

classes mapped in the field with remotely sensed data. We

tested our methodology using one Satellite Pour Lobserva-

tion de la Terre (SPOT 4) multispectral image of Para-

gominas, acquired in August 1999.

2. Background

2.1. Causes and impacts of forest degradation in the

Eastern Amazon

Forest degradation in the Amazon has been characterized

through field studies, particularly in the arc of deforestation

along the southern and eastern border of the Amazon

(Cochrane et al., 1999; Gerwing, 2002; Johns et al., 1996;

Uhl & Bushbacher, 1985). Degradation is generally grouped

into classes based on the degrading activity and intensity.

These activities include selective logging and burning, and

the quantification of the impacts is measured in terms of

forest structure and composition (Table 1).

Selective logging degrades forests during harvesting

activities, causing extensive damage to nearby trees and

soils (Johns et al., 1996), increasing the risk of species

Fig. 1. Forest degradation spatial arrangement and composition (vegetation as presented by intact forest and forest regeneration; NPV and soil) as seen from

IKONOS Geo multispectral-panchromatic data fusion product [R (0.77 0.88 Am), G (0.640.72 Am) B (0.520.61 Am) and pan (0.450.90 Am)], acquired in

November 2000 in the study area.

C. Souza Jr. et al. / Remote Sensing of Environment 87 (2003) 494506 495


3/13

extinction (Martini, Rosa, & Uhl, 1994), as well as increas-

ing carbon emissions (Houghton, 1995). In addition, timber

extraction increases fire susceptibility (Holdsworth & Uhl,

1997) and open roads, which encourage unplanned devel-

opment (Verssimo, Barreto, Tarifa, & Uhl, 1995). During

harvesting, selective logging kills or damages 1046% of

living biomass (Gerwing, 2002; Verssimo, Barreto, Mattos,

Tarifa, & Uhl, 1992; Verssimo et al., 1995), and damages

an area of 1503 2276 m2/ha (Johns et al., 1996). The

number of pioneer species can increase by 500% and the

canopy cover can decrease to 63% of the original intact

forest depending on the intensity and age of degradation

(Gerwing, 2002).

Forest burning often occurs after forests have been

opened through logging and are almost always started by

human activities in adjacent areas (e.g., slash-and-burn).

Forest fires can kill more than 45% of trees more than 20 cm

in diameter the first time a forest is burned, opening the

canopy by 3050% and increasing the susceptibility of the

forests to subsequent fires (Cochrane et al., 1999). In

addition, pioneer species increase in density by 689%.Subsequent fires can then kill up to 98% of trees more than

20 cm in diameter and leave only 1040% of the original

canopy (Cochrane et al., 1999).

2.2. Mapping forest degradation in the Eastern Amazon

using remote sensing

Prior studies in the Amazon have generally focused on

detection of one or two types of degraded forests through

remote sensing analysis. Mixture models (Adams et al.,

1995) have been pointed out as one of the most appropriate

techniques to map such degraded environments since it is

capable of separating soil, vegetation, nonphotosynthetic

vegetation (Roberts, Adams, & Smith, 1993; Roberts et al.,

1998) and shade abundance at a sub-pixel scale. This

technique enables the detection of subtle changes, and gives

physical meaning to the spectral data provided by the

satellite.

Soil fraction images, obtained through mixture models,

also allow the detection of small areas (log landings) in the

forest cleared to store timber (Monteiro, Souza, & Barreto,

2003; Souza & Barreto, 2000). Based on a site-specific

harvesting radius, it is possible to estimate the area

affected by selective logging (Souza & Barreto, 2000).

Besides its simplicity and applicability in forest surveil-

lance programs, this methodology does not provide infor-

mation on the degree of forest damage. On the other hand,

the fraction of nonphotosynthetic vegetation (NPV) may

allow estimates of forest degradation since it is associated

with dead vegetation (Roberts et al., 1993). Cochrane and

Souza (1998) used the NPV fraction image to map recently

and old burned forests in the Eastern Amazon and pointed

out its potential to map several levels of forest degradation.

3. Study area

Paragominas is a logging center in the northeast of Para,

covering 19,310 km with a population of 76,095 inhabitants

(IBGE, 2000) (Fig. 2). It was founded in 1965 after the

opening of the Belem-Brasilia highway, one of the first and

only major roads connecting north and south of Brazil.

Through the 1970s and 1980s, ranching was the dominant

land use in the area. However, the 1990s brought a boom in

the logging industry, which bought wood from the land oflarge ranches and extracted it at extremely high rates (35

40 m3 wood/ha; Verssimo et al., 1992). Since the logging

boom, the municipality has experienced a collapse of the

timber industry, damaging the local economy (Schneider,

Arima, Verssimo, Barreto, & Souza, 2002).

The natural vegetation is almost entirely dense, closed

upland forest. The dry season extends from JuneNovem-

ber, with annual temperatures averaging 25 jC. This area

was chosen because its long history of logging, ranching

and agricultural activity allowed for the inclusion of all

forest degradation types. In addition, numerous field studies

in the area allow for characterization of forest degradation

classes and determination of the history and age of forest

stands (Gerwing, 2002; Johns et al., 1996; Uhl & Bush-

bacher, 1985; Verssimo et al., 1992).

4. Methods

4.1. Data set and preprocessing

Data acquired for this study included one multispectral

SPOT 4 scene (bands: 0.500.59, 0.610.68, 0.790.89,

1.581.75 Am; 20 m pixel size) acquired in August of 1999.

Table 1

Characterization of forest degradation classes at the field scale (source: Gerwing, 2002)

Forest degradation class Physical characteristics Biomass (tons/ha) Canopy cover

Intact

forest (%)

Wood

debris (%)

Disturbed

soil (%)

Burned

vegetation (%)

Total live Total dead Percent of

original

Intact forest 99 (1) 1 (1) < 1 (0) 0 309 (28) 55 (13) 98 (0)

Moderately logged 72 (6) 11 (8) 17 (8) 0 245 (18) 76 (9) 97 (1)Heavily logged 38 (4) 32 (9) 30 (6) 0 168 (8) 149 (23) 63 (11)

Lightly burned 27 (10) 26 (3) 10 (5) 37 (11) 177 (3) 101 (37) 84 (8)

Heavily burned 2 (2) 42 (5) 13 (8) 42 (11) 50 (14) 128 (15) 39 (20)

Standard deviation in parentheses.

C. Souza Jr. et al. / Remote Sensing of Environment 87 (2003) 494506496


4/13

Fig. 3. Methods used to map forest degradation in the study area.

Fig. 2. Map of the study area showing the location of the SPOT 4 and IKONOS images.



5/13

Four Landsat Thematic Mapper (TM) images acquired in

1988 (16 August), 1991 (24 July), 1995 (05 July) and 1996

(03 June) were used to aid in the identification of forest

degradation age for the purpose of the collection of training

samples. The images cover an area of 3666 km2, mostly

within the boundaries of Paragominas County in the north-

east of Para (Fig. 2). The SPOT 4 image was geometricallycorrected using a set of 18 ground control points extracted

from topographic maps (Diretoria do Servico Geografico

DSG/IBGE). Additionally, two IKONOS fusion scenes

(11 m pixels; 15 15 km each scene), acquired in

November 2000 within the study area, were used in com-

bination with field data to identify samples to train a

decision tree classifier, and to generate test fields to evaluate

the accuracy of the final classification (Figs. 1 and 2).

4.2. Forest mask

The first step was to separate forest and nonforest pixels

(i.e., pasture, secondary growth, plantation, cloud/shadow

and water classes). We tested different methods proposed in

the literature to perform this task such as segmentation of

the shade fraction image (Shimabukuro et al., 1998), tas-

seled-cap brightness index (Cochrane & Souza, 1998) and

unsupervised classification. The ISODATA unsupervised

classification algorithm produced the best result for a

forest/nonforest map. Areas that presented spectral ambigu-

ity ( < 5%) between burned pasture and burned forest; and

forest regeneration and secondary forest, were corrected

using post-classification manual editing. Forest areas of less

than 1 ha were also automatically reclassified as nonforest

and a mean filter (5 5 pixels) was used to close forest

gaps. These procedures allowed us to isolate the forested

pixels from the other land cover classes (Fig. 3).

4.3. Mixture models

We estimated the amount of shade, soil, green vegetation

and NPV within each pixel of the SPOT 4 image. Candidate

pixels for pure spectra of these materials were found using

the Pixel Purity IndexPPI (Boardman, Kruse, & Green,

1995). The PPI results were inspected in terms of the shape

of the spectral curves and field context (e.g., soils are mostly

Fig. 4. Fraction image composition of the forest degradation classes defined at the field scale plotted in a triangular space. Vertices of the ternary diagram

(clockwise from top) represent 100% of NPV, shade and vegetation endmembers.

Table 2

Characterization of forest degradation image classes in terms of field classes

Image classification

scheme

Field description

Intact forest Consists of mature, undisturbed forests dominated

by shade tolerant species

Logged forest Recently moderate logged (up to 1 year) or old

heavily logged forest (>5 years) which haveexperienced extensive damage due to harvesting

activities, such as road creation, small clearings

for timber piles, and structural damage to nearby

trees and soil

Degraded forest Includes recently burned, heavily burned or

heavily logged forest (up to 2 years)

Forest regeneration Regeneration of old heavily burned and burned

forests



6/13

associated with roads) to define the final set of image pure

spectra, known as image endmembers. The final green

vegetation endmember was extracted from green pasture

and the shade endmember was extracted from a dark water

body. The soil endmember was extracted from road and

NPV was acquired from senesced grass found in a pasture.

We applied least-square linear mixture modeling

(Adams, Smith, & Gillespie, 1993) to estimate the propor-

tion of each endmember within the SPOT 4 pixels. The sum

of the fractions should add up to 1, but our mixture

modeling did not use this condition as a constraint. Themixture modeling equations are given by

DNb Xn

i1

FiDNi;b eb

for

Xn

i1

Fi 1

where DNb is encoded radiance in band b, Fi the fraction of

endmember i, DNi,b encoded radiance for endmember i, inband b, and eb is the residual error for each band.

The mixing model results were evaluated as proposed by

Adams et al. (1995). First, we evaluated the root-mean-

square (RMS) image. Our final model showed very low

RMS ( < 1 DN values). Next, fraction images were evaluated

and interpreted in terms of field context and spatial distri-

bution. The fraction image inspections also allowed us to

evaluate the spectral separability of NPV and soil endmem-

bers. For example, the final mixture model generated soil

Fig. 5. Binary decision tree used to classify forest degradation and intact forest classes using the fraction images derived from the SPOT 4 spectral mixture

model. The decision tree divides a data set into progressively more homogeneous classes based on a series of hierarchical binary decisions. In the decision tree

shown, the first branch splits at 22.5% NPVand branches left if this is true, right if false. Following this split, intact forest would follow the left branch with less

than 68% green vegetation and less than 15.5% NPV.

Table 3

Forest class statistics extracted from the training samples used in the DTC

algorithm

Classes Fraction

images

Minimum

(%)

Mean

(%)

Maximum

(%)

Standard

deviation

Intact forest Shade 26.0 36.6 48.0 4.5

NPV 9.0 13.6 17.0 1.9

Vegetation 39.0 49.5 60.0 4.7Logging Shade 15.0 27.8 41.0 6.2

NPV 14.0 19.9 26.0 2.9

Vegetation 43.0 52.6 66.0 5.1

Heavily logged Shade 24.0 33.8 45.0 4.7

NPV 20.0 29.3 48.0 6.0

Vegetation 19.0 37.1 54.0 7.3

Old logging Shade 12.0 25.6 36.0 4.7

NPV 15.0 18.0 22.0 1.6

Vegetation 46.0 56.5 66.0 4.6

Burned forest Shade 20.0 31.9 42.0 4.8

NPV 20.0 29.6 44.0 6.0

Vegetation 25.0 40.2 52.0 6.6

Heavily burned Shade 26.0 33.9 42.0 3.6

NPV 28.0 38.4 46.0 4.8

Vegetation 23.0 29.2 39.0 4.5

Burning Shade 0.0 11.1 32.0 8.9

regeneration NPV 15.0 18.3 22.0 1.9

Vegetation 51.0 69.8 79.0 8.2



7/13

fraction values slightly negative in forested pixel and high

fraction values along roads and bare soil values. The NPV

fraction was high in pasture and degraded forest areas.

Finally, we inspected the histograms of the fraction

images to quantify the percentage of pixels laying outside

the range of 0% to 100%. Our endmembers modeled about

95% of the pixels within the physically meaningful range.After acquiring a good model, we normalized the NPV,

green vegetation and shade fractions to fit within the range

from 0% to 100%. The normalization procedure was nec-

essary because forested pixels are composed essentially of

green vegetation, NPV and shade endmembers.

4.4. Forest degradation classification

We acquired GPS coordinates at the 14 research sites

used by Gerwing (2002) in his forest degradation charac-

terization study to identify the following classes of de-

graded forest: heavily logged forest, moderately logged

forest; heavily burned forest and burned forest (Table 1).

The forest inventories were conducted in November 1999

(Gerwing, personal communication). The GPS coordinates

were collected, in a nondifferential mode, using a 12

channel Trimble Pathfinder unit (Trimble Navigation,Chandler, AZ). These coordinates were superimposed on

the SPOT 4 image and used as a reference to select areas

(3 3 pixels) to extract fraction values of NPV, green

vegetation, and shade of each field class of degraded

forest.

We did not attempt to extract pixels from the forest

transects (10 500 m) used by Gerwing (2002) because the

forest canopy did not enable us to accurately locate the

forest transects with our GPS unit. However, the procedure

we chose to extract fraction values from the image did not

Fig. 6. Fraction images obtained from the SPOT 4 mixture model. Shade fraction is low in deforested areas and high in canopy gaps (a). Heavily logged forest

shows high NPV percent as well as areas that were logged and burned (b). Vegetation fraction can also enhance canopy gap damage areas due to logging ( c

low values). Log landings, skid trails and roads can be identified from the soil fraction images (dhigh values).



8/13

represent a problem to our analysis because the transects

were representative of the degraded forest areas (Gerwing,

2002). Additional geolocated points acquired from field

surveys (n = 17) and high-resolution IKONOS (n = 4) were

combined with the forest transect data. Classes representing

burning regeneration and old logged forest (>5 years) were

also included in the analysis.

We then used box plots, scatter matrices and ternary

diagrams to evaluate which fraction image could be used to

separate the forest degradation classes defined by Gerwing

(2002). This preliminary exploratory data analysis allowed

us to evaluate whether or not it would be possible to

distinguish the field classes defined by Gerwing using the

SPOT 4 image.

Fraction images, when plotted in a ternary diagram,

revealed that a direct linkage could not be established for

all degradation classes (Fig. 4). Heavily logged, heavily

burned and burned field classes were grouped into the forest

degradation image class. Intact forest and burning regener-ation could be mapped at the SPOT 4 scale. However, old

logging and logged forest were also combined to form the

logging image class (Table 2).

4.5. Forest degradation classification

The forest fraction images derived from the mixture

models were classified into intact forest, degraded forest,

forest regeneration and logged forest using a decision tree

classifier (DTC) (Friedl & Brodlley, 1997; Roberts et al.,

1998) available in the statistical software package S-Plus.

We used the same training samples extracted for the

exploratory data analysis (45 pixels/class), except that the

outliers were excluded before running the DTC. The final

forest degradation map was obtained after applying a

majority filter (5 5) to eliminate small isolated classes.

Fig. 7. Map of forest degradation of the study area derived from fraction images using the decision tree classification rules.

Table 4

Area of the image classes estimated from the forest degradation map

Classes Area

(km2)

Percent of

the study area

Percent of

the forest

Nonforest 1569.3 42.8

Intact forest 728.8 19.9 34.7

Degraded forest 185.6 5.1 8.9

Regeneration 11.8 0.3 0.6Logged forest 1171.1 31.9 55.8

Total 3666.6 100.0 100



9/13

4.6. Accuracy assessment analysis

In order to evaluate the classification accuracy, 300 points

were randomly generated for each of the forest degradation

classes and the nonforest class within the area covered by the

two IKONOS images (Fig. 2). Accuracy assessment for the

intact forest class was not possible because the two IKONOSimages were acquired in areas where there was no such class.

Visual inspection of IKONOS data, at a 1:10,000 scale, was

used to identify the reference class of each point within an

area equivalent to the SPOT 4 pixel (i.e., 20 20 m).

Because the IKONOS and SPOT 4 images were not acquired

on the same dates (1 year apart), we had to use a color

composite of the SPOT 4 (R4, G3, B2) to help to assign the

final class to the randomly selected points. Points were only

used if the class could be confidently identified in the

IKONOS images. We also eliminated points that fell on

cloud/shadow areas in the IKONOS images.

5. Results

5.1. Mixture models

The fraction images derived from the mixture model

provided useful information to map degraded forests (i.e.,

heavily burned, heavily logged and burned forest field

classes; Table 2, Figs. 4 and 6) in the study area. NPV

and shade fractions increased in areas that were heavily

burned or heavily logged, whereas vegetation content de-

Fig. 8. Relationship between percent of NPV, estimated from the SPOT image, and total live aboveground biomass (a) and wood debris combined with burned

vegetation (b) acquired from field inventory conducted by Gerwing (2002). Error bars represent F1 standard deviation.



10/13

creased (Fig. 6). Degraded forests showed NPV fractions

higher than 20% (Table 3, Fig. 5). This result agrees with

the threshold value found by Cochrane and Souza (1998) to

separate forest from recently burned forest in Tailandia, a

county neighboring Paragominas in the Eastern Amazon.

The heavily burned field class showed the highest NPV

fraction values (mean = 38.4%), whereas heavily loggedforest and burned forest had similar NPV mean values

(mean = 29.3% and 29.6%, respectivelyTable 3, Figs. 5

and 6). The main difference between the heavily logged

class and burned forest class is in the green vegetation

fraction. The heavily logged forests showed a lower shade

(mean = 33.8%) content than the burned forests (mean =

31.9%; Table 3).

Intact forest showed good correlation between field and

image scales. This class was characterized by NPV values

lower than 17%, a vegetation fraction varying from 39% to

60%; and shade content ranging from 26% to 48% (Table 3,

Fig. 5). Logged forest exhibited a decrease in shade

(mean = 27.8%) and a subtle increase in the NPV content

(mean = 19.9%) which allowed its separability from intact

forest (Table 3, Figs. 5 and 6). Logged forest (Table 4)

showed a higher range of green vegetation (4360%) than

intact forest because selective logging creates a heteroge-

neous environment composed of intact forest, canopy-re-

duced areas, small cleared areas, liana-rich areas, and areas

that forest regeneration took place.

Regeneration of degraded forest resulted in a strong

decrease in the shade fraction (mean = 11%) and increase

in vegetation (mean = 69.8%) when compared to intact

forest. This class is generally associated with older degraded

forest areas. Visual inspection of the Landsat TM imagesshowed that some of these areas are more than 10 years old.

Soil fractions were slightly negative in the intact forest

class and all forest degradation classes (Fig. 6). However,

important spatial information was obtained from the soil

fraction image. For example, small areas of land cleared for

temporary storage of logs in the forest (log landings) were

detected in the soil fraction (Fig. 6). This information can be

used to separate recently logged forest from old logged forest

by applying a harvesting radius from the log landings (180

m; Souza & Barreto, 2000). Using this approach we esti-

mated that 1676 ha of forest had been recently logged (about

1.5% of the total logged forest). Log landings were not

associated with old logged and degraded forest because

regeneration and canopy closure erases this spatial signature.

In addition, logging roads and skid trails were also identified

in the soil fraction images. This information can be useful to

model logging economic extent(Souza, Monteiro, Salomao,

& Valente, 2001) and map forest areas that are economically

accessible (Verssimo, Souza, Stone, & Uhl, 1997).

5.2. Forest degradation map

As the statistics in Table 4 show, a simple forest/non-

forest characterization of an area such as Paragominas

would drastically misrepresent the extent of intact forests.

Of the entire area (3666 km2), about 57% of the area would

be classified as intact forest, whereas only 20% (729 km2) is

actually intact forest when the degraded forest classes are

included as a class (Table 4, Fig. 7).

The remaining forested areas are in some stage of forest

degradation, which includes logged forest (1171 km2

; 5% ofthe forested area) and degraded forest (185.6 km2; 9% of the

forested area). Severely degraded forests are forests which

have recently been heavily burned or drastically damaged

by selective logging and have not had time to regenerate.

Less than 1% of the forested area was mapped as regener-

ation (Table 4).

Even though there was not a one-to-one relationship

between the field and image classes, we found a strong

correlation between NPV and total live aboveground

biomass and combined wood debris and burned vegeta-

tion (Fig. 8). There is a negative linear relationship

between the mean NPV fraction values and mean total

live aboveground biomass (R2 = 0.97) measured in the

field. NPV content showed a positive correlation with the

combined wood debris and burned vegetation percents

(R2 = 0.89).

5.3. Accuracy assessment

The accuracy was high for the forest/nonforests map

(93%; Table 5). However, when trying to assess the

accuracy of the forest degradation classes, the difference

in the acquisition date of the SPOT 4 image and the

IKONOS image imposed some difficulties due to regener-

ation of previously degraded areas and the appearance ofnew degraded areas. The overall classification of the forest

degradation classes was 86%. Areas that were classified as

logged forest showed 92.2% and 73.8% users and pro-

ducers accuracy, respectively. Degraded forest was satis-

factorily mapped (65.5% and 83.7% users and producers

accuracy, respectively; Table 5). No estimate is provided

for the burning regeneration class because only a few

randomly selected points (n = 6) were representative of this

class.

Table 5

Accuracy assessment results of the forest degradation map

IKONOS reference Forest degradation map Total Users

image Nonforest Degraded

forest

Logged

forest

accuracy

Nonforest 132 6 4 142 93.0

Degraded forest 2 36 17 55 65.5

Logged forest 4 1 59 64 92.2

Total 138 43 80 261

Producers accuracy 95.7 83.7 73.8

Overall accuracy = 0.86

Kappa coefficient = 0.78

Forest/nonforest accuracy=((132/142)+(113a/119))/2 = 0.93

a Degraded forest+ logged forest.



11/13

6. Discussion

6.1. Field and satellite scale linkage

We cannot compare directly the forest composition

measured at the field scale (Table 1) and at the pixel scale

(Table 3). For example, the percentage of intact forestmeasured in the field represents the amount of vegetation

that had not been disturbed by logging and/or burning, and

not the percentage of green vegetation. Additionally, the

intact forest field class is composed of leaves (green

vegetation), barks and branches. The percentage of intact

forest is high for the intact forest class and drops in

degraded forests, reaching 2% in heavily burned forests.

As a result, there is an increase in wood debris and burned

vegetation (NPV) as intact forest becomes more degraded

(Table 1).

At the satellite pixel scale, the intact forest is character-

ized by a mixture of green vegetation (mean = 49%), shade

(mean = 36%) and NPV (mean = 13%; Table 3). The highest

green vegetation fraction values is found in regeneration

class (mean = 69%), followed by old logged class

(mean= 56%) and logging class (mean= 52%; Table 3).

This pattern occurs because the forest gaps caused by

logging activity closes rapidly (Johns et al., 1996; Verssimo

et al., 1992) causing a decrease in shade content. As a result,

green vegetation fraction increases in logged forests. For

this reason, green vegetation estimated at the sub-pixel scale

should not be interpreted as a measure of disturbance of

forested areas.

The other differences observed at the field and satellite

scales are associated with NPV estimates. These differencesoccur because the natural NPV contents (i.e., bark,

branches, stems, and dead leaves) were not measured at

the field scale. The NPV reported in Table 1 refers to wood

debris and burned vegetation resulted from logging and/or

burned activities. The NPV estimated with the mixture

models represents the total NPV content (i.e., natural and

anthropogenic). The offset observed in the regression line of

wood debris and burned vegetation versus NPV fraction

(Fig. 8b) is due to the different methodologies used to

estimate NPV content in the field and in the satellite image.

6.2. Fraction image classification

Our SPOT 4 mixture model results agree with prior

studies that used fraction images, derived from Landsat

TM, to map burned forest (Cochrane & Souza, 1998), log

landings created by selective logging (Souza & Barreto,

2000), and intact forest (Adams et al., 1995; Cochrane &

Souza, 1998; Roberts et al., 1998) in the Amazon.

Prior studies have shown that a unique set of thresholds

obtained with DTC can be used to map land cover classes

locally and that the rules cannot be extended to other regions

without considering natural vegetation type variations (Rob-

erts et al., 1998, in press). The classification thresholds

obtained in our study to map classes of degraded forest must

not be interpreted as general rules to map degraded forest.

Variation on harvesting intensity, forest burning dynamics,

and logging and fire recurrence regime will also affect the

portability of the DTC thresholds. Additionally, regenera-

tion of degraded forest can contribute to change the class

thresholds. However, in areas like Paragominas, that under-went drastic forest structure and composition changes due to

logging and fire activities, the threshold might also show

temporal instability. A time series analysis of degraded

forest area is required to evaluate the temporal portability

of the DTC thresholds.

6.3. Mapping of forest degradation

The NPV endmember was the most important end

member to separate classes of degraded forests at the SPOT

4 image scale. Given the low spectral resolution of SPOT 4

and spectral resemblance with soil spectra, locating the NPV

endmember was not a straightforward process. We used

three approaches to identify the NPV endmember and

separate NPV from soil. First, we computed the PPI

algorithm to identify potential endmember candidates. Sec-

ond, we compared the PPI results with field knowledge to

identify areas more likely to find NPV (e.g., pastures with

senesced grass) and soil (associated with roads and bare

areas). Finally, we evaluated the fraction image results in

terms of expected class composition (e.g., intact forest must

have no soil and low to moderate NPV values).

The NPV fraction image improved the mapping of old

logged forest when compared to soil fraction image used by

Souza and Barreto (2000). The rapid regeneration of loglanding areas (Gerwing, 2002; Johns et al., 1996; Stone &

Lefebvre, 1998) limits the use of soil fraction images to

mapping selective logging no more than 2 years old.

Moreover, information about the harvesting radius is needed

in order to estimate the area affected by logging using soil

fractions. Additionally, soil fraction images also fail to

identify log landing in highly disturbed forest areas because

the spatial signal of the log landing is confused with other

types of cleared areas. The NPV fraction allowed us to map

selectively logged forest up to 10 years old. It is important

to highlight that most of these forests were logged at least

twice. For that reason, the NPV fraction may not improve

the mapping of very old logged forest in areas that logging

intensity is low (>20 m3/ha), such as Sinop, in Mato Grosso

state (Monteiro et al., 2003) where soil fraction images have

also proved to be effective to map selectively logged areas

up to 2 years old.

Our analysis did not include information on edge effects

and minimum size of a forest patch that would still function

as a forest. The total area of intact forests would be reduced

if information on edge effects and minimum forest area had

been included in the analysis. There are three main blocks of

intact forest in the eastern portion of the study area (Fig. 7).

These forest blocks show evidence of selective logging and



12/13

deforestation activities along the border with deforested

areas. This may lead to the conclusion that these forest

blocks are no longer pristine. As a final consideration, our

classification procedure may be improved by using images

acquired at different dates, ideally every year. Old degraded

forest (>10 years) may have regenerated and misclassified at

old logged forest. Having been used images acquired indifferent dates, disallowed class transitions could have been

defined to avoid this type of classification error.

7. Conclusions

This study showed that a partial linkage of field classes

with moderate-resolution satellite images such as SPOT 4

(20 m IFOV), and potentially Landsat TM (30 m IFOV), can

be achieved by deriving information at a sub-pixel scale

associated with field materials. Improvements can be made

to fully integrate these scales by including both temporal

and spatial information.

Our results reinforce the need for a more complete

characterization of forests in areas of the Amazon where

logging centers and agricultural activities play an important

economic role. A classification scheme that is reflective of

basin-wide degraded forest types found in the field is

needed for integration of data over time and scale. In order

to accomplish this task, more field studies must be con-

ducted covering other forest types than have been selective-

ly logged, as for example, transitional forest in Mato Grosso

(Monteiro et al., 2003). Field inventories and map of

degraded forests in the region are needed for local and state

governments to identify resource bases and effectively planconservation and development activities. Additionally, more

accurate estimates of the extent and type of degraded forests

are needed to incorporate into ecological, economic and

carbon models for the basin.

Acknowledgements

The authors thank PPG7 Programa de Pesquisa

Dirigida (MMA/MCT/FINEP) for the financial support

to the Instituto do Homem e Meio Ambiente da Amazonia-

IMAZON for conducting the forest inventory and acquiring

the SPOT 4 and Landsat TM 5 images. The work of Carlos

Souza Jr. and Luciano Moreira, and the acquisition of the

IKONOS images, were funded through the TREES II

Project (European Community-Joint Research Center). The

work of Laurel Firestone was financially supported by

Brown University. We also thank NASA/LBA through Dr.

Dar Roberts for providing laboratory resources and financ-

ing the work of Carlos Souza Jr. during the preparation of

this manuscript. Finally, we are very thankful to Dr. Jeffrey

Gerwing for sharing his forest inventory database, and

Danielle Dalsoren and Dylan Prentiss for their comments on

this manuscript.

References

Adams, J. B., Sabol, D. E., Kapos, V., Almeida Filho, R., Robert, D.,

Smith, M. O., & Gillespie, A. R. (1995). Classification of multispectral

images based on fractions of endmembers: Application to land-cover

change in the Brazilian Amazon. Remote Sensing of Environment, 52,

137154.Adams, J. B., Smith, M. O., & Gillespie, A. R. (1993). Imaging spectro-

scopy: Interpretation based on spectral mixture analysis. In V. M. Pieters,

& P. Englert (Eds.), Remote Geochemical Analysis: Elemental and Min-

eralogical Composition, vol. 7 (pp. 145166). New York: Cambridge

Univ. Press.

Boardman, J. W., Kruse, F. A., & Green, R. O. (1995). Mapping target

signatures via partial unmixing of AVIRIS data. Summaries of the 5th

Airborne Earth Science Workshop, January 2326. JPL Publication,

95(1), 23 26.

Cochrane, M. A., Alencar, A., Schulze, M., Souza Jr., C., Nepstad, D.,

Lefebvre, P., & Davidson, E. (1999). Positive feedbacks in the fire

dynamic of closed canopy tropical forest. Science, 284, 18321835.

Cochrane, M. A., & Schulze, M. D. (1999). Fire as a recurrent event in

tropical forest of the eastern Amazon: Effects on forest structure, bio-

mass, and species composition. Biotropica, 31(1), 2 16.Cochrane, M. A., & Souza Jr., C. (1998). Linear mixture model classifica-

tion of burned forest in the eastern Amazon. International Journal of

Remote Sensing, 19(17), 3433 3440.

Firestone, L. A., & Souza Jr., C. (2002). The role of remote sensing and

GIS in enforcement of areas of permanent preservation in the Brazilian

Amazon. Geocarto, 17(2), 5156.

Friedl, M. A., & Brodlley, C. E. (1997). Decision tree classification of land

cover from remotely sensed data. Remote Sensing of Environment, 61,

339409.

Gerwing, J. J. (2002). Degradation of forests through logging and fire in the

eastern Brazilian Amazon. Forest Ecology and Management, 157(1),

131141.

Holdsworth, A. R., & Uhl, C. (1997). Fire in Amazonian selectively logged

rain forest and the potential for fire reduction. Ecological Applications,

7, 713 725.Houghton, R. A. (1995). Determining emissions of carbon from land: A

global strategy. In S. Murai (Ed.), Toward global planning of sus-

tainable use of the Earth. Development of Global Eco-Engineering,

( pp. 59 75). Amsterdam: Elsevier.

IBGE (2000). Censo Demografico Brasileiro, http://www.ibge.gov.br.

INPE, Instituto Nacional de Pesquisas Espaciais (2000). Monitoring of

the Brazilian Amazonian forest by satellite 19981999, Sao Jose dos

Campos.

IPAM, WHRC, IMAZON, INPE, INMET, & NASA (1998). Mapa de

Risco de Incendios Florestais e Queimadas Agrcolas na Amazonia

Brasileira para o Segundo Semestre de 1998. PPG7-USAID, 27 pp.

Johns, J., Barreto, P., & Uhl, C. (1996). Logging damage in planned and

unplanned logging operation and its implications for sustainable timber

production in the eastern Amazon. Forest Ecology and Management,

89, 59 77.Martini, A., Rosa, N., & Uhl, C. (1994). An attempt to predict which

Amazonian tree species may be threatened by logging activities. Envi-

ronmental Conservation, 21(2), 152162.

Monteiro, A. L., Souza Jr., C., & Barreto, P. (2003). Detection of logging in

Amazonian transition forests using spectral mixture models. Interna-

tional Journal of Remote Sensing, 1(24), 151159.

Nepstad, D. C., Verssimo, J. A., Alencar, A., Nobre, C., Lima, E., Lefeb-

vre, P., Schlesinger, P., Potter, C., Moutinho, P., Mendoza, E., Cochrane,

M., & Brooks, V. (1999). Large-scale impoverishment of Amazonian

forests by logging and fire. International Weekly Journal of Science,

Nature, 398, 504 508.

Roberts, D. A., Adams, B. J., & Smith, M. O. (1993). Green vegetation,

nonphotosynthetic vegetation, and soil in AVIRIS data. Remote Sensing

of Environment, 14, 255269.

http://%20http//www.ibge.gov.brhttp://%20http//www.ibge.gov.brhttp://%20http//www.ibge.gov.br


13/13

Roberts, D. A., Batista, G. T., Pereira, J. L. G., Walker, E. K., & Nelson,

B. W. (1998). Change identification using multitemporal spectral mix-

ture analysis: Applications in eastern Amazon. In R. S. Luneetta, & C.

D. Elvidge (Eds.), Remote Sensing Change Detection: Environmental

Monitoring Methods and Applications, vol. 9 (pp. 137159). Chelsea,

MI: Ann Arbor Press.

Schneider, R. R., Arima, E., Verssimo, A., Barreto, P., & Souza Jr., C.

(2002). Sustainable Amazon: Limitations and opportunities for rural

development. World Bank Technical Paper No. 515. Environmental

Series, Washington, DC, 50 pp.

Shimabukuro, Y. E., Batista, G. T., Mello, E. M. K., Moreira, J. C., &

Duarte, V. (1998). Using shade fraction image segmentation to evaluate

deforestation in Landsat Thematic Mapper images of the Amazon Re-

gion. International Journal of Remote Sensing, 19(3), 535541.

Souza Jr., C., & Barreto, P. (2001). Sistema de fiscalizacao, licenciamento e

monitoramento de propriedades rurais de Mato Grosso. Causas e dina-

mica do desmatamento n a Amazonia, Brasil (pp. 307334). Braslia,

DF Brasil: Ministerio do Meio Ambiente. In Portuguese.

Souza Jr., C., & Barreto, P. (2000). An alternative approach for detecting

and monitoring selectively logged forests in the Amazon. International

Journal of Remote Sensing, 21(1), 173179.

Souza Jr., C., Monteiro, A. L., Salomao, R., & Valente, A. (2001). Extracao

de Informacoes de Imagens Landsat para Modelos de Alcance Econo m-

ico da Atividade Madeireira. SBSRSimposio Brasileiro de Sensoria-

mento Remoto. Foz de Iguacu/PR, 2126 de Abril de 2001. In

Portuguese.

Stone, T. A., & Lefebvre, P. A. (1998). Using multi-temporal satellite data

to evaluate selective logging in Para, Brazil. International Journal of

Remote Sensing, 13, 25172526.

Uhl, C., & Bushbacher, R. (1985). A disturbing synergism between cattle

ranch burning practices and selective tree harvesting in the eastern

Amazon. Biotropica, 17, 265 268.

Verssimo, A., Barreto, P., Mattos, M., Tarifa, R., & Uhl, C. (1992). Log-

ging impacts and prospects for sustainable forest management in an old

Amazon frontier: The case of Paragominas. Forest Ecology and Man-

agement, 55, 169199.

Verssimo, A., Barreto, P., Tarifa, R., & Uhl, C. (1995). Extraction of a

high-value natural resource from Amazon: The case of mahogany. For-

est Ecology and Management, 72, 39 60.

Verssimo, A., Souza Jr., C., Stone, S., & Uhl, C. (1997). Zoning of

timber extraction in the Brazilian Amazon. Conservation Biology,

12(1), 110.

Watrin, O. S., & Rocha, A. M. A. (1992). Levantamento da vegetac ao

natural e do uso da terra no Municpio de Paragominas (PA) utilizando

imagens TM/Landsat. Boletim de Pesquisa, vol. 124 (p. 40). Belem, PA:

EMBRAPA/CPATU. In Portuguese.


monitoreo de degradación forestal con imágenes spot4

Documents